ExLL: An Extremely Low-latency Congestion Control for Mobile Cellular Networks Shinik Park Jinsung Lee Junseon Kim UNIST University of Colorado Boulder UNIST [email protected] [email protected] [email protected] Jihoon Lee Sangtae Ha Kyunghan Lee University of Colorado Boulder University of Colorado Boulder UNIST [email protected] [email protected] [email protected] ABSTRACT 1 INTRODUCTION Since the diagnosis of severe bufferbloat in mobile cellular networks, Recent observations in the Internet revealed that TCP sessions often a number of low-latency congestion control algorithms have been suffer from exceptionally long packet delay even in a small band- proposed. However, due to the need for continuous bandwidth prob- width delay product (BDP) network. Gettys et al. termed this phe- ing in dynamic cellular channels, existing mechanisms are designed nomenon as bufferbloat [13] and diagnosed it as an over-buffering to cyclically overload the network. As a result, it is inevitable that problem at the bottleneck link, mostly caused by TCP’s loss-based their latency deviates from the smallest possible level (i.e., mini- congestion control mechanism that keeps pushing packets to the mum RTT). To tackle this problem, we propose a new low-latency network until the bottleneck queue becomes full. congestion control, ExLL, which can adapt to dynamic cellular Follow-up measurement studies [14, 18, 21] confirmed that the channels without overloading the network. To do so, we develop bufferbloat problem is often severe in cellular networks. They diag- two novel techniques that run on the cellular receiver: 1) cellular nosed that Cubic [15], the default TCP congestion control algorithm bandwidth inference from the downlink packet reception pattern in Linux (hence, the default in most Android devices) and in Win- and 2) minimum RTT calibration from the inference on the uplink dows [7], is aggressive enough to quickly fill up cellular network scheduling interval. Furthermore, we incorporate the control frame- buffers, resulting in up to several hundred milliseconds of additional work of FAST into ExLL’s cellular specific inference techniques. packet latency. Hence, ExLL can precisely control its congestion window to not overload the network unnecessarily. Our implementation of ExLL on Android smartphones demonstrates that ExLL reduces latency 1.1 Exploiting Cellular Network much closer to the minimum RTT compared to other low-latency Characteristics for Protocol Design congestion control algorithms in both static and dynamic channels In order to tackle such a latency problem in cellular networks, of LTE networks. there have been various research proposals on designing a low- latency congestion control algorithm [10, 21, 23, 24, 33, 36]. These CCS CONCEPTS approaches, however, have not fully leveraged cellular network • Networks → Network protocol design; Transport proto- specific characteristics. cols; Network performance analysis; For example, for the downlink scheduling the base station (BS) schedules downlink packets towards multiple user equipments KEYWORDS (UEs) at 1 ms granularity (a.k.a. transmission time interval, TTI), based on both the signal strength reported by each UE and the Congestion control, Cellular network, Transport protocol current traffic load [9]. This indicates that we can instantly infer the ACM Reference Format: cellular link bandwidth by observing packet reception patterns in a Shinik Park, Jinsung Lee, Junseon Kim, Jihoon Lee, Sangtae Ha, and Kyung- very short time window instead of explicitly probing the bandwidth han Lee. 2018. ExLL: An Extremely Low-latency Congestion Control for or measuring throughput for a relatively long period of time (e.g., a Mobile Cellular Networks. In CoNEXT ’18: International Conference on emerg- few RTTs). For the uplink scheduling, the BS needs to grant uplink ing Networking EXperiments and Technologies, December 4–7, 2018, Heraklion, Greece. ACM, New York, NY, USA, 13 pages. https://doi.org/10.1145/3281411. transmission eligibility for each UE, which happens at a regular 3281430 interval known as SR (scheduling request) periodicity [3]. We find that ignoring such a scheduling pattern often underestimates the Permission to make digital or hard copies of all or part of this work for personal or minimum RTT and leads congestion control algorithms to run in classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation unrealistic operating points. on the first page. Copyrights for components of this work owned by others than ACM Furthermore, the minimum RTT value is not generally affected must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, by the channel condition between UE and BS due to adaptive MCS to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. (modulation and coding scheme) selection used in LTE systems, CoNEXT ’18, December 4–7, 2018, Heraklion, Greece which approximately eliminates MAC-layer retransmissions, but © 2018 Association for Computing Machinery. also not affected by other UEs connected to the same BS since ACM ISBN 978-1-4503-6080-7/18/12...$15.00 https://doi.org/10.1145/3281411.3281430 per-UE queue is isolated. 307 CoNEXT ’18, December 4–7, 2018, Heraklion, Greece S. Park et al. 100 The understanding of the aforementioned factors has important implications on the design of an extremely low-latency congestion 80 0.9xBDP control and its best performance in throughput and delay. In par- 1.0xBDP ticular, to cope with highly dynamic channel conditions in cellular 60 1.1xBDP Cubic [15] networks, it is vital to obtain both achievable maximum throughput BBR [10] and minimum RTT as quickly and accurately as possible, so that the 40 PropRate [23] congestion control algorithm can always exploit up-to-date BDP Tput (Mbps) Verus [36] 20 Sprout [33] for low-latency control. Vegas [8] 95-th percentile 0 40 60 80 100 200 300 400 500 1.2 Gap from Ideal Latency: Existing Latency Latency (ms) Optimized Protocols Figure 1. Mean and 95-th percentile RTT against the aver- In order to check the latency performance of existing low latency age throughput of various congestion control algorithms congestion control algorithms including Cubic, we perform our own compared with that of sending a static congestion window measurement with an Android smartphone over a static LTE chan- around BDP over a real LTE network. nel as shown in Figure 1. We compare the performance (mean and 95-th percentile) of existing latency optimized congestion control Our comprehensive experiments carried out over commercial protocols along with a simple protocol that sends at a constant rate LTE networks confirm that ExLL can always achieve shorter RTT (β ×BDP). The minimum RTT and maximum throughput achievable which is much closer to the minimum RTT while retaining similar for the tested channel are about 47 ms and 90 Mbps, respectively. throughput of Cubic. More specifically, in a stationary scenario While Cubic suffers from long packet latency of 230 ms, BBR [10], where an Android smartphone is stably connected to an LTE net- PropRate [23] and Verus [36] achieve 104 ms, 71 ms, and 76 ms, work with 50 ms of its minimum RTT and 75 Mbps of its maximum respectively, with similar or much lower throughput1. Low-latency throughput, ExLL attains on average 66 ms RTT while maintain- algorithms show significant RTT suppression compared to Cubic, ing throughput of 72 Mbps; BBR and Cubic attain about 110 ms but their latency performance is still far from the ideal one char- and 261 ms RTT with about 70 Mbps and 75 Mbps. PropRate and acterized by sending the BDP variants (β = 0:9; 1:0; 1:1), which is Verus show lower throughput about 59 Mbps and 39 Mbps and to achieve about 68 ms at 90 Mbps throughput. It is intriguing that attain around 61 ms and 92 ms RTT. In a mobile scenario where a BBR and PropRate that are designed to track and utilize the BDP smartphone user moves between good (-95 dBm) and bad channels of the network are performing not as good as sending the BDP. (-125 dBm), ExLL retains around 61 ms and 45 Mbps while BBR and However, considering the overhead from cycling several modes Cubic stay around 78 ms and 395 ms with about 40 Mbps and 46 of operation to probe bandwidth and RTT, the existence of the Mbps. PropRate and Verus shows 53 ms and 63 ms but is with only performance gap is not surprising. 23 Mbps and 34 Mbps. In summary, our contributions are three-fold. 1.3 ExLL Contributions • We develop novel techniques that can estimate the cellular To bridge the gap, we propose a new low-latency congestion control link bandwidth and realistic minimum RTT without explicit for mobile cellular networks, namely ExLL (Extremely Low Latency) probing, which can be easily extended to next-generation that reduces latency as close to the minimum RTT while retaining cellular technologies such as 5G. the same level of throughput that Cubic achieves. To obtain such • We incorporate the control logic of FAST into ExLL to mini- performance, instead of probing the network, ExLL estimates the mize the latency even in dynamic cellular channel conditions. bandwidth of cellular channels by analyzing the packet reception • We implement ExLL in both receiver- and sender-side ver- pattern at an LTE subframe granularity in the downlink and also sions that give wider deployment opportunities. The receiver- estimates the minimum RTT more realistically by incorporating side ExLL can provide an immediate solution for untouched SR periodicity in the uplink. As these estimations can be done re- commodity servers while the sender-side ExLL can provide liably at each UE, ExLL takes the receiver-side design as its first a fundamental solution for 5G URLLC (ultra reliability and choice.